Beyond distance: modelling cyclists' route choice using calibrated empirical travel time

Abstract

This paper develops a cyclist route choice modelling framework based on calibrated empirical travel time rather than distance alone or conventional infrastructure proxies. Using a large crowdsourced GPS dataset from Copenhagen, Denmark, we identify cruising segments in smoothed and mapmatched trajectories, estimate non-parametric empirical link-level cruising speed distributions, and regularise them through hierarchical Bayesian shrinkage. We then estimate link-level delays as the difference between observed traversal time and cyclist-specific expected cruising time, again using shrinkage to ensure network-wide coverage. These cyclist- and link-specific travel times are aggregated to the route level and introduced into multinomial logit route choice models. The results show that travel-time-based specifications substantially outperform conventional distancebased models, and that the best-performing linear specification distinguishes between cruising time and delay. The framework provides a basis for future extensions toward more fundamental route costs, including, e.g., physical effort.